Research On Medical Ultrasonic Image Super-Resolution Reconstruction Method

Ultrasonic imaging is widely used in clinical diagnosis due to its noninvasive,convenient and fast.However,due to the low image contrast,there are a variety of artifacts and noise interference,and there is a high clinical experience of doctors.Therefore,it is important to improve the quality of ultrasonic image quality using image super resolution techniques.Compared to natural scene images,ultrasound images lack a pair of high and low resolution image data sets,making reconstruction tasks more challenging.In this paper,the characteristics of ultrasound images are used to study the ultrasound image of the super-resolution reconstruction problem of the generation against the network,and the nonsupervised learning.Us-Case and CCA-US,in public data sets,A large number of experiments have been carried out on the data set accumulated in the Affiliated Hospital of Qingdao University.This paper first analyzes two phases to generate a Two-Stage Ga N and the non-supervised learning model ZSSR("Zero-shot" super-resolution),based on this,two-stage ultrasound image super resolution loop generation confrontation network model(TWO-Stage ZSSR CYCLE GAN),the model has the following three points for TWO-Stage GAN: 1)Modify the Ga N network of the Two-Stage GAN network model to the Cycle GAN network based on the ZSSR model,using ultrasound image itself Super-resolution reconstruction makes the model more in line with the characteristics of the ultrasound image lack of true high-resolution images;2)Optimize the generator of the GAN network based on the multi-scale generator,extract multi-scale characteristics from the input image.The superresolution SR image under different scales is reconstructed by different scales;3)Using cyclic consistency loss,take full use of low resolution-super resolution-low resolution(LR-HR-LR)And high resolution-low resolution-super resolution(HR-LR-HR)cyclic consistency loss,to promote the generator to produce a sensible resolution result.Then,the U-NET network structure of the Two-Stage GAN first stage is optimized,by introducing residual sub-structural and attention mechanisms,while avoiding the loss of spatial information while deepening the network,using image context information extraction more Low frequency information,enhance the visual reconstruction effect.Finally,a large number of experiments on private data sets and two public data sets are performed,and the rationality and effectiveness of improved strategies are verified.Based on the research results,the ultrasound image superresolution reconstruction prototype system is designed and implemented,realtime super-resolution reconstruction.